Root/mm/page-writeback.c

1/*
2 * mm/page-writeback.c
3 *
4 * Copyright (C) 2002, Linus Torvalds.
5 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
6 *
7 * Contains functions related to writing back dirty pages at the
8 * address_space level.
9 *
10 * 10Apr2002 Andrew Morton
11 * Initial version
12 */
13
14#include <linux/kernel.h>
15#include <linux/module.h>
16#include <linux/spinlock.h>
17#include <linux/fs.h>
18#include <linux/mm.h>
19#include <linux/swap.h>
20#include <linux/slab.h>
21#include <linux/pagemap.h>
22#include <linux/writeback.h>
23#include <linux/init.h>
24#include <linux/backing-dev.h>
25#include <linux/task_io_accounting_ops.h>
26#include <linux/blkdev.h>
27#include <linux/mpage.h>
28#include <linux/rmap.h>
29#include <linux/percpu.h>
30#include <linux/notifier.h>
31#include <linux/smp.h>
32#include <linux/sysctl.h>
33#include <linux/cpu.h>
34#include <linux/syscalls.h>
35#include <linux/buffer_head.h>
36#include <linux/pagevec.h>
37
38/*
39 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
40 * will look to see if it needs to force writeback or throttling.
41 */
42static long ratelimit_pages = 32;
43
44/*
45 * When balance_dirty_pages decides that the caller needs to perform some
46 * non-background writeback, this is how many pages it will attempt to write.
47 * It should be somewhat larger than dirtied pages to ensure that reasonably
48 * large amounts of I/O are submitted.
49 */
50static inline long sync_writeback_pages(unsigned long dirtied)
51{
52    if (dirtied < ratelimit_pages)
53        dirtied = ratelimit_pages;
54
55    return dirtied + dirtied / 2;
56}
57
58/* The following parameters are exported via /proc/sys/vm */
59
60/*
61 * Start background writeback (via writeback threads) at this percentage
62 */
63int dirty_background_ratio = 10;
64
65/*
66 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
67 * dirty_background_ratio * the amount of dirtyable memory
68 */
69unsigned long dirty_background_bytes;
70
71/*
72 * free highmem will not be subtracted from the total free memory
73 * for calculating free ratios if vm_highmem_is_dirtyable is true
74 */
75int vm_highmem_is_dirtyable;
76
77/*
78 * The generator of dirty data starts writeback at this percentage
79 */
80int vm_dirty_ratio = 20;
81
82/*
83 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
84 * vm_dirty_ratio * the amount of dirtyable memory
85 */
86unsigned long vm_dirty_bytes;
87
88/*
89 * The interval between `kupdate'-style writebacks
90 */
91unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
92
93/*
94 * The longest time for which data is allowed to remain dirty
95 */
96unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
97
98/*
99 * Flag that makes the machine dump writes/reads and block dirtyings.
100 */
101int block_dump;
102
103/*
104 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
105 * a full sync is triggered after this time elapses without any disk activity.
106 */
107int laptop_mode;
108
109EXPORT_SYMBOL(laptop_mode);
110
111/* End of sysctl-exported parameters */
112
113
114/*
115 * Scale the writeback cache size proportional to the relative writeout speeds.
116 *
117 * We do this by keeping a floating proportion between BDIs, based on page
118 * writeback completions [end_page_writeback()]. Those devices that write out
119 * pages fastest will get the larger share, while the slower will get a smaller
120 * share.
121 *
122 * We use page writeout completions because we are interested in getting rid of
123 * dirty pages. Having them written out is the primary goal.
124 *
125 * We introduce a concept of time, a period over which we measure these events,
126 * because demand can/will vary over time. The length of this period itself is
127 * measured in page writeback completions.
128 *
129 */
130static struct prop_descriptor vm_completions;
131static struct prop_descriptor vm_dirties;
132
133/*
134 * couple the period to the dirty_ratio:
135 *
136 * period/2 ~ roundup_pow_of_two(dirty limit)
137 */
138static int calc_period_shift(void)
139{
140    unsigned long dirty_total;
141
142    if (vm_dirty_bytes)
143        dirty_total = vm_dirty_bytes / PAGE_SIZE;
144    else
145        dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) /
146                100;
147    return 2 + ilog2(dirty_total - 1);
148}
149
150/*
151 * update the period when the dirty threshold changes.
152 */
153static void update_completion_period(void)
154{
155    int shift = calc_period_shift();
156    prop_change_shift(&vm_completions, shift);
157    prop_change_shift(&vm_dirties, shift);
158}
159
160int dirty_background_ratio_handler(struct ctl_table *table, int write,
161        void __user *buffer, size_t *lenp,
162        loff_t *ppos)
163{
164    int ret;
165
166    ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
167    if (ret == 0 && write)
168        dirty_background_bytes = 0;
169    return ret;
170}
171
172int dirty_background_bytes_handler(struct ctl_table *table, int write,
173        void __user *buffer, size_t *lenp,
174        loff_t *ppos)
175{
176    int ret;
177
178    ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
179    if (ret == 0 && write)
180        dirty_background_ratio = 0;
181    return ret;
182}
183
184int dirty_ratio_handler(struct ctl_table *table, int write,
185        void __user *buffer, size_t *lenp,
186        loff_t *ppos)
187{
188    int old_ratio = vm_dirty_ratio;
189    int ret;
190
191    ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
192    if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
193        update_completion_period();
194        vm_dirty_bytes = 0;
195    }
196    return ret;
197}
198
199
200int dirty_bytes_handler(struct ctl_table *table, int write,
201        void __user *buffer, size_t *lenp,
202        loff_t *ppos)
203{
204    unsigned long old_bytes = vm_dirty_bytes;
205    int ret;
206
207    ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
208    if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
209        update_completion_period();
210        vm_dirty_ratio = 0;
211    }
212    return ret;
213}
214
215/*
216 * Increment the BDI's writeout completion count and the global writeout
217 * completion count. Called from test_clear_page_writeback().
218 */
219static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
220{
221    __prop_inc_percpu_max(&vm_completions, &bdi->completions,
222                  bdi->max_prop_frac);
223}
224
225void bdi_writeout_inc(struct backing_dev_info *bdi)
226{
227    unsigned long flags;
228
229    local_irq_save(flags);
230    __bdi_writeout_inc(bdi);
231    local_irq_restore(flags);
232}
233EXPORT_SYMBOL_GPL(bdi_writeout_inc);
234
235void task_dirty_inc(struct task_struct *tsk)
236{
237    prop_inc_single(&vm_dirties, &tsk->dirties);
238}
239
240/*
241 * Obtain an accurate fraction of the BDI's portion.
242 */
243static void bdi_writeout_fraction(struct backing_dev_info *bdi,
244        long *numerator, long *denominator)
245{
246    if (bdi_cap_writeback_dirty(bdi)) {
247        prop_fraction_percpu(&vm_completions, &bdi->completions,
248                numerator, denominator);
249    } else {
250        *numerator = 0;
251        *denominator = 1;
252    }
253}
254
255/*
256 * Clip the earned share of dirty pages to that which is actually available.
257 * This avoids exceeding the total dirty_limit when the floating averages
258 * fluctuate too quickly.
259 */
260static void clip_bdi_dirty_limit(struct backing_dev_info *bdi,
261        unsigned long dirty, unsigned long *pbdi_dirty)
262{
263    unsigned long avail_dirty;
264
265    avail_dirty = global_page_state(NR_FILE_DIRTY) +
266         global_page_state(NR_WRITEBACK) +
267         global_page_state(NR_UNSTABLE_NFS) +
268         global_page_state(NR_WRITEBACK_TEMP);
269
270    if (avail_dirty < dirty)
271        avail_dirty = dirty - avail_dirty;
272    else
273        avail_dirty = 0;
274
275    avail_dirty += bdi_stat(bdi, BDI_RECLAIMABLE) +
276        bdi_stat(bdi, BDI_WRITEBACK);
277
278    *pbdi_dirty = min(*pbdi_dirty, avail_dirty);
279}
280
281static inline void task_dirties_fraction(struct task_struct *tsk,
282        long *numerator, long *denominator)
283{
284    prop_fraction_single(&vm_dirties, &tsk->dirties,
285                numerator, denominator);
286}
287
288/*
289 * scale the dirty limit
290 *
291 * task specific dirty limit:
292 *
293 * dirty -= (dirty/8) * p_{t}
294 */
295static void task_dirty_limit(struct task_struct *tsk, unsigned long *pdirty)
296{
297    long numerator, denominator;
298    unsigned long dirty = *pdirty;
299    u64 inv = dirty >> 3;
300
301    task_dirties_fraction(tsk, &numerator, &denominator);
302    inv *= numerator;
303    do_div(inv, denominator);
304
305    dirty -= inv;
306    if (dirty < *pdirty/2)
307        dirty = *pdirty/2;
308
309    *pdirty = dirty;
310}
311
312/*
313 *
314 */
315static unsigned int bdi_min_ratio;
316
317int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
318{
319    int ret = 0;
320
321    spin_lock_bh(&bdi_lock);
322    if (min_ratio > bdi->max_ratio) {
323        ret = -EINVAL;
324    } else {
325        min_ratio -= bdi->min_ratio;
326        if (bdi_min_ratio + min_ratio < 100) {
327            bdi_min_ratio += min_ratio;
328            bdi->min_ratio += min_ratio;
329        } else {
330            ret = -EINVAL;
331        }
332    }
333    spin_unlock_bh(&bdi_lock);
334
335    return ret;
336}
337
338int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
339{
340    int ret = 0;
341
342    if (max_ratio > 100)
343        return -EINVAL;
344
345    spin_lock_bh(&bdi_lock);
346    if (bdi->min_ratio > max_ratio) {
347        ret = -EINVAL;
348    } else {
349        bdi->max_ratio = max_ratio;
350        bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
351    }
352    spin_unlock_bh(&bdi_lock);
353
354    return ret;
355}
356EXPORT_SYMBOL(bdi_set_max_ratio);
357
358/*
359 * Work out the current dirty-memory clamping and background writeout
360 * thresholds.
361 *
362 * The main aim here is to lower them aggressively if there is a lot of mapped
363 * memory around. To avoid stressing page reclaim with lots of unreclaimable
364 * pages. It is better to clamp down on writers than to start swapping, and
365 * performing lots of scanning.
366 *
367 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
368 *
369 * We don't permit the clamping level to fall below 5% - that is getting rather
370 * excessive.
371 *
372 * We make sure that the background writeout level is below the adjusted
373 * clamping level.
374 */
375
376static unsigned long highmem_dirtyable_memory(unsigned long total)
377{
378#ifdef CONFIG_HIGHMEM
379    int node;
380    unsigned long x = 0;
381
382    for_each_node_state(node, N_HIGH_MEMORY) {
383        struct zone *z =
384            &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
385
386        x += zone_page_state(z, NR_FREE_PAGES) +
387             zone_reclaimable_pages(z);
388    }
389    /*
390     * Make sure that the number of highmem pages is never larger
391     * than the number of the total dirtyable memory. This can only
392     * occur in very strange VM situations but we want to make sure
393     * that this does not occur.
394     */
395    return min(x, total);
396#else
397    return 0;
398#endif
399}
400
401/**
402 * determine_dirtyable_memory - amount of memory that may be used
403 *
404 * Returns the numebr of pages that can currently be freed and used
405 * by the kernel for direct mappings.
406 */
407unsigned long determine_dirtyable_memory(void)
408{
409    unsigned long x;
410
411    x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages();
412
413    if (!vm_highmem_is_dirtyable)
414        x -= highmem_dirtyable_memory(x);
415
416    return x + 1; /* Ensure that we never return 0 */
417}
418
419void
420get_dirty_limits(unsigned long *pbackground, unsigned long *pdirty,
421         unsigned long *pbdi_dirty, struct backing_dev_info *bdi)
422{
423    unsigned long background;
424    unsigned long dirty;
425    unsigned long available_memory = determine_dirtyable_memory();
426    struct task_struct *tsk;
427
428    if (vm_dirty_bytes)
429        dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
430    else {
431        int dirty_ratio;
432
433        dirty_ratio = vm_dirty_ratio;
434        if (dirty_ratio < 5)
435            dirty_ratio = 5;
436        dirty = (dirty_ratio * available_memory) / 100;
437    }
438
439    if (dirty_background_bytes)
440        background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
441    else
442        background = (dirty_background_ratio * available_memory) / 100;
443
444    if (background >= dirty)
445        background = dirty / 2;
446    tsk = current;
447    if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
448        background += background / 4;
449        dirty += dirty / 4;
450    }
451    *pbackground = background;
452    *pdirty = dirty;
453
454    if (bdi) {
455        u64 bdi_dirty;
456        long numerator, denominator;
457
458        /*
459         * Calculate this BDI's share of the dirty ratio.
460         */
461        bdi_writeout_fraction(bdi, &numerator, &denominator);
462
463        bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
464        bdi_dirty *= numerator;
465        do_div(bdi_dirty, denominator);
466        bdi_dirty += (dirty * bdi->min_ratio) / 100;
467        if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
468            bdi_dirty = dirty * bdi->max_ratio / 100;
469
470        *pbdi_dirty = bdi_dirty;
471        clip_bdi_dirty_limit(bdi, dirty, pbdi_dirty);
472        task_dirty_limit(current, pbdi_dirty);
473    }
474}
475
476/*
477 * balance_dirty_pages() must be called by processes which are generating dirty
478 * data. It looks at the number of dirty pages in the machine and will force
479 * the caller to perform writeback if the system is over `vm_dirty_ratio'.
480 * If we're over `background_thresh' then the writeback threads are woken to
481 * perform some writeout.
482 */
483static void balance_dirty_pages(struct address_space *mapping,
484                unsigned long write_chunk)
485{
486    long nr_reclaimable, bdi_nr_reclaimable;
487    long nr_writeback, bdi_nr_writeback;
488    unsigned long background_thresh;
489    unsigned long dirty_thresh;
490    unsigned long bdi_thresh;
491    unsigned long pages_written = 0;
492    unsigned long pause = 1;
493
494    struct backing_dev_info *bdi = mapping->backing_dev_info;
495
496    for (;;) {
497        struct writeback_control wbc = {
498            .sync_mode = WB_SYNC_NONE,
499            .older_than_this = NULL,
500            .nr_to_write = write_chunk,
501            .range_cyclic = 1,
502        };
503
504        get_dirty_limits(&background_thresh, &dirty_thresh,
505                &bdi_thresh, bdi);
506
507        nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
508                    global_page_state(NR_UNSTABLE_NFS);
509        nr_writeback = global_page_state(NR_WRITEBACK);
510
511        bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
512        bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
513
514        if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
515            break;
516
517        /*
518         * Throttle it only when the background writeback cannot
519         * catch-up. This avoids (excessively) small writeouts
520         * when the bdi limits are ramping up.
521         */
522        if (nr_reclaimable + nr_writeback <
523                (background_thresh + dirty_thresh) / 2)
524            break;
525
526        if (!bdi->dirty_exceeded)
527            bdi->dirty_exceeded = 1;
528
529        /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
530         * Unstable writes are a feature of certain networked
531         * filesystems (i.e. NFS) in which data may have been
532         * written to the server's write cache, but has not yet
533         * been flushed to permanent storage.
534         * Only move pages to writeback if this bdi is over its
535         * threshold otherwise wait until the disk writes catch
536         * up.
537         */
538        if (bdi_nr_reclaimable > bdi_thresh) {
539            writeback_inodes_wb(&bdi->wb, &wbc);
540            pages_written += write_chunk - wbc.nr_to_write;
541            get_dirty_limits(&background_thresh, &dirty_thresh,
542                       &bdi_thresh, bdi);
543        }
544
545        /*
546         * In order to avoid the stacked BDI deadlock we need
547         * to ensure we accurately count the 'dirty' pages when
548         * the threshold is low.
549         *
550         * Otherwise it would be possible to get thresh+n pages
551         * reported dirty, even though there are thresh-m pages
552         * actually dirty; with m+n sitting in the percpu
553         * deltas.
554         */
555        if (bdi_thresh < 2*bdi_stat_error(bdi)) {
556            bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
557            bdi_nr_writeback = bdi_stat_sum(bdi, BDI_WRITEBACK);
558        } else if (bdi_nr_reclaimable) {
559            bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
560            bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
561        }
562
563        if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
564            break;
565        if (pages_written >= write_chunk)
566            break; /* We've done our duty */
567
568        __set_current_state(TASK_INTERRUPTIBLE);
569        io_schedule_timeout(pause);
570
571        /*
572         * Increase the delay for each loop, up to our previous
573         * default of taking a 100ms nap.
574         */
575        pause <<= 1;
576        if (pause > HZ / 10)
577            pause = HZ / 10;
578    }
579
580    if (bdi_nr_reclaimable + bdi_nr_writeback < bdi_thresh &&
581            bdi->dirty_exceeded)
582        bdi->dirty_exceeded = 0;
583
584    if (writeback_in_progress(bdi))
585        return;
586
587    /*
588     * In laptop mode, we wait until hitting the higher threshold before
589     * starting background writeout, and then write out all the way down
590     * to the lower threshold. So slow writers cause minimal disk activity.
591     *
592     * In normal mode, we start background writeout at the lower
593     * background_thresh, to keep the amount of dirty memory low.
594     */
595    if ((laptop_mode && pages_written) ||
596        (!laptop_mode && ((global_page_state(NR_FILE_DIRTY)
597                   + global_page_state(NR_UNSTABLE_NFS))
598                      > background_thresh)))
599        bdi_start_background_writeback(bdi);
600}
601
602void set_page_dirty_balance(struct page *page, int page_mkwrite)
603{
604    if (set_page_dirty(page) || page_mkwrite) {
605        struct address_space *mapping = page_mapping(page);
606
607        if (mapping)
608            balance_dirty_pages_ratelimited(mapping);
609    }
610}
611
612static DEFINE_PER_CPU(unsigned long, bdp_ratelimits) = 0;
613
614/**
615 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
616 * @mapping: address_space which was dirtied
617 * @nr_pages_dirtied: number of pages which the caller has just dirtied
618 *
619 * Processes which are dirtying memory should call in here once for each page
620 * which was newly dirtied. The function will periodically check the system's
621 * dirty state and will initiate writeback if needed.
622 *
623 * On really big machines, get_writeback_state is expensive, so try to avoid
624 * calling it too often (ratelimiting). But once we're over the dirty memory
625 * limit we decrease the ratelimiting by a lot, to prevent individual processes
626 * from overshooting the limit by (ratelimit_pages) each.
627 */
628void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
629                    unsigned long nr_pages_dirtied)
630{
631    unsigned long ratelimit;
632    unsigned long *p;
633
634    ratelimit = ratelimit_pages;
635    if (mapping->backing_dev_info->dirty_exceeded)
636        ratelimit = 8;
637
638    /*
639     * Check the rate limiting. Also, we do not want to throttle real-time
640     * tasks in balance_dirty_pages(). Period.
641     */
642    preempt_disable();
643    p = &__get_cpu_var(bdp_ratelimits);
644    *p += nr_pages_dirtied;
645    if (unlikely(*p >= ratelimit)) {
646        ratelimit = sync_writeback_pages(*p);
647        *p = 0;
648        preempt_enable();
649        balance_dirty_pages(mapping, ratelimit);
650        return;
651    }
652    preempt_enable();
653}
654EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
655
656void throttle_vm_writeout(gfp_t gfp_mask)
657{
658    unsigned long background_thresh;
659    unsigned long dirty_thresh;
660
661        for ( ; ; ) {
662        get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
663
664                /*
665                 * Boost the allowable dirty threshold a bit for page
666                 * allocators so they don't get DoS'ed by heavy writers
667                 */
668                dirty_thresh += dirty_thresh / 10; /* wheeee... */
669
670                if (global_page_state(NR_UNSTABLE_NFS) +
671            global_page_state(NR_WRITEBACK) <= dirty_thresh)
672                            break;
673                congestion_wait(BLK_RW_ASYNC, HZ/10);
674
675        /*
676         * The caller might hold locks which can prevent IO completion
677         * or progress in the filesystem. So we cannot just sit here
678         * waiting for IO to complete.
679         */
680        if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
681            break;
682        }
683}
684
685/*
686 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
687 */
688int dirty_writeback_centisecs_handler(ctl_table *table, int write,
689    void __user *buffer, size_t *length, loff_t *ppos)
690{
691    proc_dointvec(table, write, buffer, length, ppos);
692    bdi_arm_supers_timer();
693    return 0;
694}
695
696#ifdef CONFIG_BLOCK
697void laptop_mode_timer_fn(unsigned long data)
698{
699    struct request_queue *q = (struct request_queue *)data;
700    int nr_pages = global_page_state(NR_FILE_DIRTY) +
701        global_page_state(NR_UNSTABLE_NFS);
702
703    /*
704     * We want to write everything out, not just down to the dirty
705     * threshold
706     */
707    if (bdi_has_dirty_io(&q->backing_dev_info))
708        bdi_start_writeback(&q->backing_dev_info, nr_pages);
709}
710
711/*
712 * We've spun up the disk and we're in laptop mode: schedule writeback
713 * of all dirty data a few seconds from now. If the flush is already scheduled
714 * then push it back - the user is still using the disk.
715 */
716void laptop_io_completion(struct backing_dev_info *info)
717{
718    mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
719}
720
721/*
722 * We're in laptop mode and we've just synced. The sync's writes will have
723 * caused another writeback to be scheduled by laptop_io_completion.
724 * Nothing needs to be written back anymore, so we unschedule the writeback.
725 */
726void laptop_sync_completion(void)
727{
728    struct backing_dev_info *bdi;
729
730    rcu_read_lock();
731
732    list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
733        del_timer(&bdi->laptop_mode_wb_timer);
734
735    rcu_read_unlock();
736}
737#endif
738
739/*
740 * If ratelimit_pages is too high then we can get into dirty-data overload
741 * if a large number of processes all perform writes at the same time.
742 * If it is too low then SMP machines will call the (expensive)
743 * get_writeback_state too often.
744 *
745 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
746 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
747 * thresholds before writeback cuts in.
748 *
749 * But the limit should not be set too high. Because it also controls the
750 * amount of memory which the balance_dirty_pages() caller has to write back.
751 * If this is too large then the caller will block on the IO queue all the
752 * time. So limit it to four megabytes - the balance_dirty_pages() caller
753 * will write six megabyte chunks, max.
754 */
755
756void writeback_set_ratelimit(void)
757{
758    ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
759    if (ratelimit_pages < 16)
760        ratelimit_pages = 16;
761    if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
762        ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
763}
764
765static int __cpuinit
766ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
767{
768    writeback_set_ratelimit();
769    return NOTIFY_DONE;
770}
771
772static struct notifier_block __cpuinitdata ratelimit_nb = {
773    .notifier_call = ratelimit_handler,
774    .next = NULL,
775};
776
777/*
778 * Called early on to tune the page writeback dirty limits.
779 *
780 * We used to scale dirty pages according to how total memory
781 * related to pages that could be allocated for buffers (by
782 * comparing nr_free_buffer_pages() to vm_total_pages.
783 *
784 * However, that was when we used "dirty_ratio" to scale with
785 * all memory, and we don't do that any more. "dirty_ratio"
786 * is now applied to total non-HIGHPAGE memory (by subtracting
787 * totalhigh_pages from vm_total_pages), and as such we can't
788 * get into the old insane situation any more where we had
789 * large amounts of dirty pages compared to a small amount of
790 * non-HIGHMEM memory.
791 *
792 * But we might still want to scale the dirty_ratio by how
793 * much memory the box has..
794 */
795void __init page_writeback_init(void)
796{
797    int shift;
798
799    writeback_set_ratelimit();
800    register_cpu_notifier(&ratelimit_nb);
801
802    shift = calc_period_shift();
803    prop_descriptor_init(&vm_completions, shift);
804    prop_descriptor_init(&vm_dirties, shift);
805}
806
807/**
808 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
809 * @mapping: address space structure to write
810 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
811 * @writepage: function called for each page
812 * @data: data passed to writepage function
813 *
814 * If a page is already under I/O, write_cache_pages() skips it, even
815 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
816 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
817 * and msync() need to guarantee that all the data which was dirty at the time
818 * the call was made get new I/O started against them. If wbc->sync_mode is
819 * WB_SYNC_ALL then we were called for data integrity and we must wait for
820 * existing IO to complete.
821 */
822int write_cache_pages(struct address_space *mapping,
823              struct writeback_control *wbc, writepage_t writepage,
824              void *data)
825{
826    int ret = 0;
827    int done = 0;
828    struct pagevec pvec;
829    int nr_pages;
830    pgoff_t uninitialized_var(writeback_index);
831    pgoff_t index;
832    pgoff_t end; /* Inclusive */
833    pgoff_t done_index;
834    int cycled;
835    int range_whole = 0;
836
837    pagevec_init(&pvec, 0);
838    if (wbc->range_cyclic) {
839        writeback_index = mapping->writeback_index; /* prev offset */
840        index = writeback_index;
841        if (index == 0)
842            cycled = 1;
843        else
844            cycled = 0;
845        end = -1;
846    } else {
847        index = wbc->range_start >> PAGE_CACHE_SHIFT;
848        end = wbc->range_end >> PAGE_CACHE_SHIFT;
849        if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
850            range_whole = 1;
851        cycled = 1; /* ignore range_cyclic tests */
852
853        /*
854         * If this is a data integrity sync, cap the writeback to the
855         * current end of file. Any extension to the file that occurs
856         * after this is a new write and we don't need to write those
857         * pages out to fulfil our data integrity requirements. If we
858         * try to write them out, we can get stuck in this scan until
859         * the concurrent writer stops adding dirty pages and extending
860         * EOF.
861         */
862        if (wbc->sync_mode == WB_SYNC_ALL &&
863            wbc->range_end == LLONG_MAX) {
864            end = i_size_read(mapping->host) >> PAGE_CACHE_SHIFT;
865        }
866    }
867
868retry:
869    done_index = index;
870    while (!done && (index <= end)) {
871        int i;
872
873        nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
874                  PAGECACHE_TAG_DIRTY,
875                  min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
876        if (nr_pages == 0)
877            break;
878
879        for (i = 0; i < nr_pages; i++) {
880            struct page *page = pvec.pages[i];
881
882            /*
883             * At this point, the page may be truncated or
884             * invalidated (changing page->mapping to NULL), or
885             * even swizzled back from swapper_space to tmpfs file
886             * mapping. However, page->index will not change
887             * because we have a reference on the page.
888             */
889            if (page->index > end) {
890                /*
891                 * can't be range_cyclic (1st pass) because
892                 * end == -1 in that case.
893                 */
894                done = 1;
895                break;
896            }
897
898            done_index = page->index + 1;
899
900            lock_page(page);
901
902            /*
903             * Page truncated or invalidated. We can freely skip it
904             * then, even for data integrity operations: the page
905             * has disappeared concurrently, so there could be no
906             * real expectation of this data interity operation
907             * even if there is now a new, dirty page at the same
908             * pagecache address.
909             */
910            if (unlikely(page->mapping != mapping)) {
911continue_unlock:
912                unlock_page(page);
913                continue;
914            }
915
916            if (!PageDirty(page)) {
917                /* someone wrote it for us */
918                goto continue_unlock;
919            }
920
921            if (PageWriteback(page)) {
922                if (wbc->sync_mode != WB_SYNC_NONE)
923                    wait_on_page_writeback(page);
924                else
925                    goto continue_unlock;
926            }
927
928            BUG_ON(PageWriteback(page));
929            if (!clear_page_dirty_for_io(page))
930                goto continue_unlock;
931
932            ret = (*writepage)(page, wbc, data);
933            if (unlikely(ret)) {
934                if (ret == AOP_WRITEPAGE_ACTIVATE) {
935                    unlock_page(page);
936                    ret = 0;
937                } else {
938                    /*
939                     * done_index is set past this page,
940                     * so media errors will not choke
941                     * background writeout for the entire
942                     * file. This has consequences for
943                     * range_cyclic semantics (ie. it may
944                     * not be suitable for data integrity
945                     * writeout).
946                     */
947                    done = 1;
948                    break;
949                }
950            }
951
952            if (wbc->nr_to_write > 0) {
953                if (--wbc->nr_to_write == 0 &&
954                    wbc->sync_mode == WB_SYNC_NONE) {
955                    /*
956                     * We stop writing back only if we are
957                     * not doing integrity sync. In case of
958                     * integrity sync we have to keep going
959                     * because someone may be concurrently
960                     * dirtying pages, and we might have
961                     * synced a lot of newly appeared dirty
962                     * pages, but have not synced all of the
963                     * old dirty pages.
964                     */
965                    done = 1;
966                    break;
967                }
968            }
969        }
970        pagevec_release(&pvec);
971        cond_resched();
972    }
973    if (!cycled && !done) {
974        /*
975         * range_cyclic:
976         * We hit the last page and there is more work to be done: wrap
977         * back to the start of the file
978         */
979        cycled = 1;
980        index = 0;
981        end = writeback_index - 1;
982        goto retry;
983    }
984    if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
985        mapping->writeback_index = done_index;
986
987    return ret;
988}
989EXPORT_SYMBOL(write_cache_pages);
990
991/*
992 * Function used by generic_writepages to call the real writepage
993 * function and set the mapping flags on error
994 */
995static int __writepage(struct page *page, struct writeback_control *wbc,
996               void *data)
997{
998    struct address_space *mapping = data;
999    int ret = mapping->a_ops->writepage(page, wbc);
1000    mapping_set_error(mapping, ret);
1001    return ret;
1002}
1003
1004/**
1005 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1006 * @mapping: address space structure to write
1007 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1008 *
1009 * This is a library function, which implements the writepages()
1010 * address_space_operation.
1011 */
1012int generic_writepages(struct address_space *mapping,
1013               struct writeback_control *wbc)
1014{
1015    /* deal with chardevs and other special file */
1016    if (!mapping->a_ops->writepage)
1017        return 0;
1018
1019    return write_cache_pages(mapping, wbc, __writepage, mapping);
1020}
1021
1022EXPORT_SYMBOL(generic_writepages);
1023
1024int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1025{
1026    int ret;
1027
1028    if (wbc->nr_to_write <= 0)
1029        return 0;
1030    if (mapping->a_ops->writepages)
1031        ret = mapping->a_ops->writepages(mapping, wbc);
1032    else
1033        ret = generic_writepages(mapping, wbc);
1034    return ret;
1035}
1036
1037/**
1038 * write_one_page - write out a single page and optionally wait on I/O
1039 * @page: the page to write
1040 * @wait: if true, wait on writeout
1041 *
1042 * The page must be locked by the caller and will be unlocked upon return.
1043 *
1044 * write_one_page() returns a negative error code if I/O failed.
1045 */
1046int write_one_page(struct page *page, int wait)
1047{
1048    struct address_space *mapping = page->mapping;
1049    int ret = 0;
1050    struct writeback_control wbc = {
1051        .sync_mode = WB_SYNC_ALL,
1052        .nr_to_write = 1,
1053    };
1054
1055    BUG_ON(!PageLocked(page));
1056
1057    if (wait)
1058        wait_on_page_writeback(page);
1059
1060    if (clear_page_dirty_for_io(page)) {
1061        page_cache_get(page);
1062        ret = mapping->a_ops->writepage(page, &wbc);
1063        if (ret == 0 && wait) {
1064            wait_on_page_writeback(page);
1065            if (PageError(page))
1066                ret = -EIO;
1067        }
1068        page_cache_release(page);
1069    } else {
1070        unlock_page(page);
1071    }
1072    return ret;
1073}
1074EXPORT_SYMBOL(write_one_page);
1075
1076/*
1077 * For address_spaces which do not use buffers nor write back.
1078 */
1079int __set_page_dirty_no_writeback(struct page *page)
1080{
1081    if (!PageDirty(page))
1082        SetPageDirty(page);
1083    return 0;
1084}
1085
1086/*
1087 * Helper function for set_page_dirty family.
1088 * NOTE: This relies on being atomic wrt interrupts.
1089 */
1090void account_page_dirtied(struct page *page, struct address_space *mapping)
1091{
1092    if (mapping_cap_account_dirty(mapping)) {
1093        __inc_zone_page_state(page, NR_FILE_DIRTY);
1094        __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1095        task_dirty_inc(current);
1096        task_io_account_write(PAGE_CACHE_SIZE);
1097    }
1098}
1099
1100/*
1101 * For address_spaces which do not use buffers. Just tag the page as dirty in
1102 * its radix tree.
1103 *
1104 * This is also used when a single buffer is being dirtied: we want to set the
1105 * page dirty in that case, but not all the buffers. This is a "bottom-up"
1106 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1107 *
1108 * Most callers have locked the page, which pins the address_space in memory.
1109 * But zap_pte_range() does not lock the page, however in that case the
1110 * mapping is pinned by the vma's ->vm_file reference.
1111 *
1112 * We take care to handle the case where the page was truncated from the
1113 * mapping by re-checking page_mapping() inside tree_lock.
1114 */
1115int __set_page_dirty_nobuffers(struct page *page)
1116{
1117    if (!TestSetPageDirty(page)) {
1118        struct address_space *mapping = page_mapping(page);
1119        struct address_space *mapping2;
1120
1121        if (!mapping)
1122            return 1;
1123
1124        spin_lock_irq(&mapping->tree_lock);
1125        mapping2 = page_mapping(page);
1126        if (mapping2) { /* Race with truncate? */
1127            BUG_ON(mapping2 != mapping);
1128            WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1129            account_page_dirtied(page, mapping);
1130            radix_tree_tag_set(&mapping->page_tree,
1131                page_index(page), PAGECACHE_TAG_DIRTY);
1132        }
1133        spin_unlock_irq(&mapping->tree_lock);
1134        if (mapping->host) {
1135            /* !PageAnon && !swapper_space */
1136            __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1137        }
1138        return 1;
1139    }
1140    return 0;
1141}
1142EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1143
1144/*
1145 * When a writepage implementation decides that it doesn't want to write this
1146 * page for some reason, it should redirty the locked page via
1147 * redirty_page_for_writepage() and it should then unlock the page and return 0
1148 */
1149int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1150{
1151    wbc->pages_skipped++;
1152    return __set_page_dirty_nobuffers(page);
1153}
1154EXPORT_SYMBOL(redirty_page_for_writepage);
1155
1156/*
1157 * Dirty a page.
1158 *
1159 * For pages with a mapping this should be done under the page lock
1160 * for the benefit of asynchronous memory errors who prefer a consistent
1161 * dirty state. This rule can be broken in some special cases,
1162 * but should be better not to.
1163 *
1164 * If the mapping doesn't provide a set_page_dirty a_op, then
1165 * just fall through and assume that it wants buffer_heads.
1166 */
1167int set_page_dirty(struct page *page)
1168{
1169    struct address_space *mapping = page_mapping(page);
1170
1171    if (likely(mapping)) {
1172        int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
1173#ifdef CONFIG_BLOCK
1174        if (!spd)
1175            spd = __set_page_dirty_buffers;
1176#endif
1177        return (*spd)(page);
1178    }
1179    if (!PageDirty(page)) {
1180        if (!TestSetPageDirty(page))
1181            return 1;
1182    }
1183    return 0;
1184}
1185EXPORT_SYMBOL(set_page_dirty);
1186
1187/*
1188 * set_page_dirty() is racy if the caller has no reference against
1189 * page->mapping->host, and if the page is unlocked. This is because another
1190 * CPU could truncate the page off the mapping and then free the mapping.
1191 *
1192 * Usually, the page _is_ locked, or the caller is a user-space process which
1193 * holds a reference on the inode by having an open file.
1194 *
1195 * In other cases, the page should be locked before running set_page_dirty().
1196 */
1197int set_page_dirty_lock(struct page *page)
1198{
1199    int ret;
1200
1201    lock_page_nosync(page);
1202    ret = set_page_dirty(page);
1203    unlock_page(page);
1204    return ret;
1205}
1206EXPORT_SYMBOL(set_page_dirty_lock);
1207
1208/*
1209 * Clear a page's dirty flag, while caring for dirty memory accounting.
1210 * Returns true if the page was previously dirty.
1211 *
1212 * This is for preparing to put the page under writeout. We leave the page
1213 * tagged as dirty in the radix tree so that a concurrent write-for-sync
1214 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
1215 * implementation will run either set_page_writeback() or set_page_dirty(),
1216 * at which stage we bring the page's dirty flag and radix-tree dirty tag
1217 * back into sync.
1218 *
1219 * This incoherency between the page's dirty flag and radix-tree tag is
1220 * unfortunate, but it only exists while the page is locked.
1221 */
1222int clear_page_dirty_for_io(struct page *page)
1223{
1224    struct address_space *mapping = page_mapping(page);
1225
1226    BUG_ON(!PageLocked(page));
1227
1228    ClearPageReclaim(page);
1229    if (mapping && mapping_cap_account_dirty(mapping)) {
1230        /*
1231         * Yes, Virginia, this is indeed insane.
1232         *
1233         * We use this sequence to make sure that
1234         * (a) we account for dirty stats properly
1235         * (b) we tell the low-level filesystem to
1236         * mark the whole page dirty if it was
1237         * dirty in a pagetable. Only to then
1238         * (c) clean the page again and return 1 to
1239         * cause the writeback.
1240         *
1241         * This way we avoid all nasty races with the
1242         * dirty bit in multiple places and clearing
1243         * them concurrently from different threads.
1244         *
1245         * Note! Normally the "set_page_dirty(page)"
1246         * has no effect on the actual dirty bit - since
1247         * that will already usually be set. But we
1248         * need the side effects, and it can help us
1249         * avoid races.
1250         *
1251         * We basically use the page "master dirty bit"
1252         * as a serialization point for all the different
1253         * threads doing their things.
1254         */
1255        if (page_mkclean(page))
1256            set_page_dirty(page);
1257        /*
1258         * We carefully synchronise fault handlers against
1259         * installing a dirty pte and marking the page dirty
1260         * at this point. We do this by having them hold the
1261         * page lock at some point after installing their
1262         * pte, but before marking the page dirty.
1263         * Pages are always locked coming in here, so we get
1264         * the desired exclusion. See mm/memory.c:do_wp_page()
1265         * for more comments.
1266         */
1267        if (TestClearPageDirty(page)) {
1268            dec_zone_page_state(page, NR_FILE_DIRTY);
1269            dec_bdi_stat(mapping->backing_dev_info,
1270                    BDI_RECLAIMABLE);
1271            return 1;
1272        }
1273        return 0;
1274    }
1275    return TestClearPageDirty(page);
1276}
1277EXPORT_SYMBOL(clear_page_dirty_for_io);
1278
1279int test_clear_page_writeback(struct page *page)
1280{
1281    struct address_space *mapping = page_mapping(page);
1282    int ret;
1283
1284    if (mapping) {
1285        struct backing_dev_info *bdi = mapping->backing_dev_info;
1286        unsigned long flags;
1287
1288        spin_lock_irqsave(&mapping->tree_lock, flags);
1289        ret = TestClearPageWriteback(page);
1290        if (ret) {
1291            radix_tree_tag_clear(&mapping->page_tree,
1292                        page_index(page),
1293                        PAGECACHE_TAG_WRITEBACK);
1294            if (bdi_cap_account_writeback(bdi)) {
1295                __dec_bdi_stat(bdi, BDI_WRITEBACK);
1296                __bdi_writeout_inc(bdi);
1297            }
1298        }
1299        spin_unlock_irqrestore(&mapping->tree_lock, flags);
1300    } else {
1301        ret = TestClearPageWriteback(page);
1302    }
1303    if (ret)
1304        dec_zone_page_state(page, NR_WRITEBACK);
1305    return ret;
1306}
1307
1308int test_set_page_writeback(struct page *page)
1309{
1310    struct address_space *mapping = page_mapping(page);
1311    int ret;
1312
1313    if (mapping) {
1314        struct backing_dev_info *bdi = mapping->backing_dev_info;
1315        unsigned long flags;
1316
1317        spin_lock_irqsave(&mapping->tree_lock, flags);
1318        ret = TestSetPageWriteback(page);
1319        if (!ret) {
1320            radix_tree_tag_set(&mapping->page_tree,
1321                        page_index(page),
1322                        PAGECACHE_TAG_WRITEBACK);
1323            if (bdi_cap_account_writeback(bdi))
1324                __inc_bdi_stat(bdi, BDI_WRITEBACK);
1325        }
1326        if (!PageDirty(page))
1327            radix_tree_tag_clear(&mapping->page_tree,
1328                        page_index(page),
1329                        PAGECACHE_TAG_DIRTY);
1330        spin_unlock_irqrestore(&mapping->tree_lock, flags);
1331    } else {
1332        ret = TestSetPageWriteback(page);
1333    }
1334    if (!ret)
1335        inc_zone_page_state(page, NR_WRITEBACK);
1336    return ret;
1337
1338}
1339EXPORT_SYMBOL(test_set_page_writeback);
1340
1341/*
1342 * Return true if any of the pages in the mapping are marked with the
1343 * passed tag.
1344 */
1345int mapping_tagged(struct address_space *mapping, int tag)
1346{
1347    int ret;
1348    rcu_read_lock();
1349    ret = radix_tree_tagged(&mapping->page_tree, tag);
1350    rcu_read_unlock();
1351    return ret;
1352}
1353EXPORT_SYMBOL(mapping_tagged);
1354

Archive Download this file



interactive